Deep Multi-Label Multi-Instance Classification on 12-Lead ECG

Abstract

As part of the PhysioNet/Computing in Cardiology Challenge 2020, we developed an end-to-end deep neural network model based on 1D ResNet and an attention-based multi-instance classification (MIC) mechanism, named as MIC-ResNet, requiring minimal signal preprocessing, for identifying 27 cardiac abnormalities from 12-lead ECG data. Our team, ECGLearner, achieved a challenge validation score of 0.486 and a full test score of 0.001, placing us 33 out of 41 in the official ranking of this year's challenge.